This project focuses on a novel framework for modelling
human-robot interactions. It is expected to result in
significant improvements in the accuracy of models of
human motion, with a minimal increase in computational
burden over existing methods.
Specifically, the current work investigates the modelling
of the behaviour of people within a dynamic environment.
It is well known that changing aspects of the environment
will influence the way people move within it. The correlations
between the behaviour of people and the state of the
environment have not, however, been investigated deeply.
The project adopts an interest-based approach to investigate
both the gross motion of people through a space and
the effects which controllable aspects of the environment
will have on this motion.
The primary modelling technique being applied is Dynamic
Bayesian Networks (DBNs), a probabilistic modelling
technique with scope for great flexibility. DBNs require
training, and the application of intelligent training
techniques to these systems allows the learning of human
behaviours from existing data. The Fish-Bird project
is being used as a data-collection tool to facilitate
this work, and over 7 Gb of log files have been collected
of (anonymous) human motion through the installation
A diagram showing 2 consecutive slices
of a DBN. The arrows show the probabilistic relationships
beteween the observed (grey) and hidden (white) variables.
These relationships occur both within a time slice,
and also between two time-slices, allowing estimation
of the hidden states and prediction of future states
given the present and past states.